Multi-Class Segmentation Network Based on Tumor Tissue in Endometrial Cancer Pathology Images: ECMTrans-net

Endometrial cancer has the second highest incidence of malignant tumors in the female reproductive system. Accurate and efficient analysis of endometrial cancer pathology images is one of the important research components of computer-aided diagnosis. However, endometrial cancer pathology images have challenges such as smaller solid tumors, lesion areas varying in morphology, and difficulty distinguishing solid and nonsolid tumors, which would affect the accuracy of subsequent pathologic analyses. An Endometrial Cancer Multi-class Transformer Network (ECMTrans-net) is therefore proposed herein to improve the segmentation accuracy of endometrial cancer pathology images. An ECM-Attention module can sequentially infer attention maps along three separate dimensions (channel, local spatial, and global spatial) and multiply the attention maps and the input feature map for adaptive feature refinement. This approach may solve the problems of the small size of solid tumors and similar characteristics of solid tumors to nonsolid tumors and further improve the accuracy of segmentation of solid tumors. In addition, an ECM-Transformer module is proposed, which can fuse multi-class feature information and dynamically adjust the receptive field, solving the issue of complex tumor features. Experiments on the Solid Tumor Endometrial Cancer Pathological (ST-ECP) data set found that performance of the ECMTrans-net was superior to state-of-the-art image segmentation methods, and the average values of accuracy, Mean Intersection over Union, precision, and Dice coefficients were 0.952, 0.927, 0.931, and 0.901, respectively.